One prompt, three dialects

The running example

Take one prompt from the library:

a vintage tram at dusk, blue hour, long exposure

A subject, a time, and a camera treatment. Each word comes from a wall: vehicles, weather, styles. Now watch it change shape three times.

The flags dialect: Midjourney

Midjourney takes a comma phrase followed by flags. Flags start with two dashes and control the frame, not the content.

/imagine prompt: a vintage tram at dusk, blue hour, long exposure --ar 3:2

The rules that matter:

• Word order is weight. Early words steer harder than late words. Put the subject first.
• --ar sets the shape. 3:2 for scenes, 2:3 for portraits, 16:9 for wide shots.
• --no is the eraser. --no people, text removes what you keep getting and don't want.
• Keep the phrase under about forty words. Midjourney reads everything but trusts the start.

The full flag list lives on the Midjourney page.

The settings dialect: Stable Diffusion

Stable Diffusion splits the ask across two boxes and a settings panel. The prompt box takes the same comma list. The negative box takes everything you want gone. The sliders replace flags.

a vintage tram at dusk, blue hour, long exposure, sharp focus

Negative promptblurry, watermark, extra wheels, deformed

The rules that matter:

• The negative prompt is standard practice. A short stock list fixes most artifacts. The negative prompts guide has tested sets.
• Aspect ratio is a resolution setting, picked in the UI, not written in the prompt.
• Quality words still earn their place. Terms like sharp focus and detailed do more here than in the other two dialects.
• Checkpoint choice changes the vocabulary. A photoreal model and an anime model read the same words into different pictures.

The sentence dialect: chat models

Nano Banana, GPT-Image, and the other chat image models read language, not keyword soup. Ask the way you would ask a person who happens to paint fast.

Make a 3:2 image: a vintage tram at dusk in blue hour light, shot as a long exposure so the lights streak.

The rules that matter:

• Sentences beat lists. The model parses grammar, so relationships survive: "a cat on the tram" stays on the tram.
• The conversation is the tool. Follow up with "same scene, but raining" instead of rewriting from scratch.
• Name the shape in words. Say wide or square or portrait, or give the ratio in the sentence.
• Wall words still work. They are compact ways to name a look, and the chat models know them all. The LLM image prompts guide goes deeper.

Translating between dialects

The translation is mechanical, which is why the converter on the tools page can do it for you:

• Flags → settings: drop the slash command, move --ar into the resolution picker, move --no into the negative box.
• Lists → sentences: pick the subject, make it the object of "Make an image of", and fold the look words into clauses.
• Sentences → lists: strip the grammar. Keep the nouns, the look words, and the light.

What never translates: Midjourney flags pasted into Stable Diffusion sit in the prompt as dead text, and negative prompts pasted into a chat model read as a request to include those things. Translate the intent, never the syntax.

Which dialect, when

• Midjourney when you want strong opinions and fast beauty with little tuning.
• Stable Diffusion when you want control: models, negatives, seeds, and repeatable results.
• A chat model when the picture needs editing by conversation, readable text, or follows from something you are already discussing.

The words stay the same in all three. Learn the vocabulary once on the walls, then spend it in whichever dialect is open in front of you.

Related walls

Published here first.